Who Says What to Whom on Twitter

Shaomei Wu

Jake M. Hofman

Cornell University, USA

Yahoo! Research, NY, USA

sw475@cornell.edu
Winter A. Mason

hofman@yahoo-inc.com

Yahoo! Research, NY, USA

Yahoo! Research, NY, USA

winteram@yahooinc.com

ABSTRACT
We study several longstanding questions in media communications research, in the context of the microblogging service
Twitter, regarding the production, ow, and consumption of
information. To do so, we exploit a recently introduced feature of Twitter known as lists to distinguish between elite
usersby which we mean celebrities, bloggers, and representatives of media outlets and other formal organizationsand
ordinary users. Based on this classication, we nd a striking concentration of attention on Twitter, in that roughly
50% of URLs consumed are generated by just 20K elite
users, where the media produces the most information, but
celebrities are the most followed. We also nd signicant
homophily within categories: celebrities listen to celebrities,
while bloggers listen to bloggers etc; however, bloggers in
general rebroadcast more information than the other categories. Next we re-examine the classical two-step ow theory of communications, nding considerable support for it
on Twitter. Third, we nd that URLs broadcast by dierent
categories of users or containing dierent types of content
exhibit systematically dierent lifespans. And nally, we examine the attention paid by the dierent user categories to
dierent news topics.

Categories and Subject Descriptors

H.1.2 [Models and Principles]: User/Machine Systems;
J.4 [Social and Behavioral Sciences]: Sociology

General Terms
two-step ow, communications, classication

Keywords
Communication networks, Twitter, information ow

1. INTRODUCTION
A longstanding objective of media communications research is encapsulated by what is known as Lasswells maxim:

Part of this research was performed while the author was

visiting Yahoo! Research, New York. The author was also
supported by NSF grant IIS-0910664.

Copyright is held by the International World Wide Web Conference Committee (IW3C2). Distribution of these papers is limited to classroom use,
and personal use by others.
WWW 2011, March 28April 1, 2011, Hyderabad, India.
ACM 978-1-4503-0637-9/11/03.

Duncan J. Watts
djw@yahoo-inc.com

who says what to whom in what channel with what effect [12], so-named for one of the pioneers of the eld,
Harold Lasswell. Although simple to state, Laswells maxim
has proven dicult to answer in the more-than 60 years
since he stated it, in part because it is generally dicult to
observe information ows in large populations, and in part
because dierent channels have very dierent attributes and
eects. As a result, theories of communications have tended
to focus either on mass communication, dened as oneway message transmissions from one source to a large, relatively undierentiated and anonymous audience, or on interpersonal communication, meaning a two-way message
exchange between two or more individuals. [16].
Correspondingly, debates among communication theorists
have tended to revolve around the relative importance of
these two putative modes of communication. For example, whereas early theories such as the hypodermic needle
model posited that mass media exerted direct and relatively
strong eects on public opinion, mid-century researchers [13,
9, 14, 4] argued that the mass media inuenced the public only indirectly, via what they called a two-step ow of
communications, where the critical intermediate layer was
occupied by a category of media-savvy individuals called
opinion leaders. The resulting limited eects paradigm
was then subsequently challenged by a new generation of
researchers [6], who claimed that the real importance of the
mass media lay in its ability to set the agenda of public
discourse. But in recent years rising public skepticism of
mass media, along with changes in media and communication technology, have tilted conventional academic wisdom
once more in favor of interpersonal communication, which
some identify as a new era of minimal eects [2].
Recent changes in technology, however, have increasingly
undermined the validity of the mass vs. interpersonal dichotomy itself. On the one hand, over the past few decades
mass communication has experienced a proliferation of new
channels, including cable television, satellite radio, specialist book and magazine publishers, and of course an array
of web-based media such as sponsored blogs, online communities, and social news sites. Correspondingly, the traditional mass audience once associated with, say, network
television has fragmented into many smaller audiences, each
of which increasingly selects the information to which it is
exposed, and in some cases generates the information itself [15]. Meanwhile, in the opposite direction interpersonal
communication has become increasingly amplied through
personal blogs, email lists, and social networking sites to

aord individuals ever-larger audiences. Together, these

two trends have greatly obscured the historical distinction
between mass and interpersonal communications, leading
some scholars to refer instead to masspersonal communications [16].
A striking illustration of this erosion of traditional media categories is provided by the micro-blogging platform
Twitter. For example, the top ten most-followed users on
Twitter are not corporations or media organizations, but
individual people, mostly celebrities. Moreover, these individuals communicate directly with their millions of followers
via their tweets, often managed by themselves or publicists,
thus bypassing the traditional intermediation of the mass
media between celebrities and fans. Next, in addition to
conventional celebrities, a new class of semi-public individuals like bloggers, authors, journalists, and subject matter
experts has come to occupy an important niche on Twitter, in some cases becoming more prominent (at least in
terms of number of followers) than traditional public gures
such as entertainers and elected ocials. Third, in spite
of these shifts away from centralized media power, media
organizationsalong with corporations, governments, and
NGOsall remain well represented among highly followed
users, and are often extremely active. And nally, Twitter
is primarily made up of many millions of users who seem
to be ordinary individuals communicating with their friends
and acquaintances in a manner largely consistent with traditional notions of interpersonal communication.
Twitter, therefore, represents the full spectrum of communications from personal and private to masspersonal to traditional mass media. Consequently it provides an interesting
context in which to address Lasswells maxim, especially as
Twitterunlike television, radio, and print mediaenables
one to easily observe information ows among the members
of its ecosystem. Unfortunately, however, the kinds of effects that are of most interest to communications theorists,
such as changes in behavior, attitudes, etc., remain dicult
to measure on Twitter. Therefore in this paper we limit
our focus to the who says what to whom part of Laswells
maxim.
To this end, our paper makes three main contributions:
We introduce a method for classifying users using Twitter Lists into elite and ordinary users, further classifying elite users into one of four categories of interest
media, celebrities, organizations, and bloggers.
We investigate the ow of information among these
categories, nding that although audience attention is
highly concentrated on a minority of elite users, much
of the information they produce reaches the masses
indirectly via a large population of intermediaries.
We nd that dierent categories of users emphasize different types of content, and that dierent content types
exhibit dramatically dierent characteristic lifespans,
ranging from less than a day to months.
The remainder of the paper proceeds as follows. In the
next section, we review related work. In Section 3 we discuss our data and methods, including Section 3.3 in which
we describe how we use Twitter Lists to classify users, outline two dierent sampling methods, and show that they
deliver qualitatively similar results. In Section 4 we analyze the production of information on Twitter, particularly

who pays attention to whom. In section 4.1, we revisit the

theory of the two-step owarguably the dominant theory
of communications for much of the past 50 yearsnding
considerable support for the theory. In Section 5, we consider who listens to what, examining rst who shares what
kinds of media content, and second the lifespan of URLs as a
function of their origin and their content. Finally, in Section
6 we conclude with a brief discussion of future work.

2.

RELATED WORK

Aside from the communications literature surveyed above,

a number of recent papers have examined information diffusion on Twitter. Kwak et al. [11] studied the topological
features of the Twitter follower graph, concluding from the
highly skewed nature of the distribution of followers and the
low rate of reciprocated ties that Twitter more closely resembled an information sharing network than a social network
a conclusion that is consistent with our own view. In addition, Kwak et al. compared three dierent measures of
inuencenumber of followers, page-rank, and number of
retweetsnding that the ranking of the most inuential
users diered depending on the measure. In a similar vein,
Cha et al. [3] compared three measures of inuencenumber
of followers, number of retweets, and number of mentions
and also found that the most followed users did not necessarily score highest on the other measures. Weng et al. [17]
compared number of followers and page rank with a modied
page-rank measure which accounted for topic, again nding
that ranking depended on the inuence measure. Finally,
Bakshy et al. [1] studied the distribution of retweet cascades
on Twitter, nding that although users with large follower
counts and past success in triggering cascades were on average more likely to trigger large cascades in the future, these
features are in general poor predictors of future cascade size.
Our paper diers from this earlier work by shifting attention from the ranking of individual users in terms of various
inuence measures to the ow of information among dierent categories of users. In this sense, it is related to recent
work by Crane and Sornette [5], who posited a mathematical model of social inuence to account for observed temporal patterns in the popularity of YouTube videos, and also
to Gomez et al [7], who studied the diusion of information among blogs and online news sources. Here, however,
our focus is on identifying specic categories of elite users,
who we dierentiate from ordinary users in terms of their
visibility, and understanding their role in introducing information into Twitter, as well as how information originating
from traditional media sources reaches the masses.

3.
3.1

DATA AND METHODS

Twitter Follower Graph

In order to understand how information is transmitted on

Twitter, we need to know the channels by which it ows;
that is, who is following whom on Twitter. To this end, we
used the follower graph studied by Kwak et al. [11], which
included 42M users and 1.5B edges. This data represents
a crawl of the graph seeded with all users on Twitter as
observed by July 31st, 2009, and is publicly available1 . As
reported by Kwak et al. [11], the follower graph is a directed
1
The
data
is
free
to
download
http://an.kaist.ac.kr/traces/WWW2010.html

from

network characterized by highly skewed distributions both of

in-degree (# followers) and out-degree (# friends, Twitter
nomenclature for how many others a user follows); however,
the out-degree distribution is even more skewed than the
in-degree distribution. In both friend and follower distributions, for example, the median is less than 100, but the maximum # friends is several hundred thousand, while a small
number of users have millions of followers. In addition, the
follower graph is also characterized by extremely low reciprocity (roughly 20%)in particular, the most-followed individuals typically do not follow many others. The Twitter
follower graph, in other words, does not conform to the usual
characteristics of social networks, which exhibit much higher
reciprocity and far less skewed degree distributions [10], but
instead resembles more the mixture of one-way mass communications and reciprocated interpersonal communications
described above.

3.2 Twitter Firehose

In addition to the follower graph, we are interested in the
content being shared on Twitter, and so we examined the
corpus of all 5B tweets generated over a 223 day period from
July 28, 2009 to March 8, 2010 using data from the Twitter
rehose, the complete stream of all tweets2 . Because our
objective is to understand the ow of information, it is useful for us to restrict attention to tweets containing URLs,
for two reasons. First, URLs add easily identiable tags to
individual tweets, allowing us to observe when a particular
piece of content is either retweeted or subsequently reintroduced by another user. And second, because URLs point
to online content outside of Twitter, they provide a much
richer source of variation than is possible in the typical 140
character tweet 3 . Finally, we note that almost all URLs
broadcast on Twitter have been shortened using one of a
number of URL shorteners, of which the most popular is
http://bit.ly/. From the total of 5B tweets recorded during
our observation period, therefore, we focus our attention on
the subset of 260M containing bit.ly URLs; thus all subsequent counts are implicitly understood to be restricted to
this content.

3.3 Twitter Lists

Our method for classifying users exploits a relatively recent feature of Twitter: Twitter Lists. Since its launch on
November 2, 2009, Twitter Lists have been used extensively
to group sets of users into topical or other categories, and
thereby to better organize and/or lter incoming tweets. To
create a Twitter List, a user provides a name (required) and
description (optional) for the list, and decides whether the
new list is public (anyone can view and subscribe to this list)
or private (only the list creator can view or subscribe to this
list). Once a list is created, the user can add/edit/delete
list members. As the purpose of Twitter Lists is to help
users organize users they follow, the name of the list can
be considered a meaningful label for the listed users. The
2

http://dev.twitter.com/doc/get/statuses/rehose
Naturally, this restriction also has downsides, in particular
that some users may be more likely to include URLs in their
tweets than others, and thus will appear to be relatively
more active and/or have more impact than if we were instead
to consider all tweets. For our purposes, however, we believe
that the practical advantages of the restriction outweigh the
potential for bias.
3

classication of users can therefore eectively exploit the

wisdom of crowds with these created lists, both in terms
of their importance to the community (number of lists on
which they appear), and also how they are perceived (e.g.
news organization vs. celebrity, etc.).
Before describing our methods for classifying users in terms
of the lists on which they appear, we emphasize that we
are motivated by a particular set of substantive questions
arising out of communications theory. In particular, we
are interested in the relative importance of mass communications, as practiced by media and other formal organizations, masspersonal communications as practiced by celebrities and prominent bloggers, and interpersonal communications, as practiced by ordinary individuals communicating
with their friends. In addition, we are interested in the relationships between these categories of users, motivated by
theoretical arguments such as the theory of the two-step
ow [9]. Rather than pursuing a strategy of automatic classication, therefore, our approach depends on dening and
identifying certain predetermined classes of theoretical interest, where both approaches have advantages and disadvantages. In particular, we restrict our attention to four
classes of what we call elite users: media, celebrities, organizations, and bloggers, as well as the relationships between
these elite users and the much larger population of ordinary users.
Analytically, our approach has some disadvantages. In
particular, by determining the categories of interest in advance, we reduce the possibility of discovering unanticipated
categories that may be of equal or greater relevance than
those we selected. Thus although we believe that for our particular purposes, the advantages of our approachnamely
conceptual clarity and ease of interpretationoutweigh the
disadvantages, automated classication methods remain an
interesting topic for future work. Finally, in addition to
these theoretically-imposed constraints, our proposed classication method must also satisfy a practical constraint
namely that the rate limits established by Twitters API
eectively preclude crawling all lists for all Twitter users4 .
Thus we instead devised two dierent sampling schemesa
snowball sample and an activity sampleeach with some
advantages and disadvantages, discussed below.

3.3.1

Snowball sample of Twitter Lists

The rst method for identifying elite users employed snowball sampling. For each category, we chose a number u0 of
seed users that were highly representative of the desired category and appeared on many category-related lists. For each
of the four categories above, the following seeds were chosen:
Celebrities: Barack Obama, Lady Gaga, Paris Hilton
Media: CNN, New York Times
Organizations: Amnesty International, World Wildlife
Foundation, Yahoo! Inc., Whole Foods
4
The Twitter API allows only 20K calls per hour, where at
most 20 lists can be retrieved for each API call. Under the
modest assumption of 40M users, where each user is included
on at most 20 lists, this would require roughly 11 weeks.
Clearly this time could be reduced by deploying multiple
accounts, but it also likely underestimates the real time quite
signicantly, as many users appear on many more than 20
lists (e.g. Lady Gaga appears on nearly 140,000).

 Blogs5 : BoingBoing, FamousBloggers, problogger, mashable. Chrisbrogan, virtuosoblogger, Gizmodo, Ileane,

dragonblogger, bbrian017, hishaman, copyblogger, engadget, danielscocco, BlazingMinds, bloggersblog, TycoonBlogger, shoemoney, wchingya, extremejohn,
GrowMap, kikolani, smartbloggerz, Element321, brandonacox, remarkablogger, jsinkeywest, seosmarty, NotAProBlog, kbloemendaal, JimiJones, ditesco
After reviewing the lists associated with these seeds, the
following keywords were hand-selected based on (a) their
representativeness of the desired categories; and (b) their
lack of overlap between categories:
Celebrities: star, stars, hollywood, celebs, celebrity,
celebrities, celebsveried, celebrity-list,celebrities-ontwitter, celebrity-tweets
Media: news, media, news-media
Organizations: company, companies, organization,
organisation, organizations, organisations, corporation,
brands, products, charity, charities, causes, cause, ngo

u0
l0
u1
l1
u2
l2
Figure 1:
Method

Schematic of the Snowball Sampling

Table 1: Distribution of users over categories

category
celeb
media
org
blog
total

Snowball Sample
# of users
% of users
82,770
15.8%
216,010
41.2%
97,853
18.7%
127,483
24.3%
524,116
100%

Activity Sample
# of users
% of users
14,778
13.0%
40,186
35.3%
14,891
13.1%
43,830
38.6%
113,685
100%

Blogs: blog, blogs, blogger, bloggers

Having selected the seeds and the keywords for each category, we then performed a snowball sample of the bipartite
graph of users and lists (see Figure 1). For each seed, we
crawled all lists on which that seed appeared. The resulting
list of lists was then pruned to contain only the l0 lists
whose names matched at least one of the chosen keywords
for that category. For instance, Lady Gaga is on lists called
faves, celebs, and celebrity, but only the latter two lists
would be kept after pruning. We then crawled all u1 users
appearing in the pruned list of lists (for instance, nding all users that appeared in the celebrity list with Lady
Gaga), and then repeated these last two steps to complete
the crawl. In total, 524, 116 users were obtained, who appeared on 7, 000, 000 lists; however, many of the more prominent users appeared on lists in more than one categoryfor
example Oprah Winfrey was frequently included in lists of
celebrity as well as media. To resolve this ambiguity, we
computed a user is membership score in category c:
nic
wic =
,
Nc
where nic is the number of lists in category c that contain
user i and Nc is the total number of lists in category c.
We then assigned each user to the category in which he
or she had the highest membership score (i.e., belonged to
the highest fraction of the categorys lists). The number of
users assigned in this manner to each category is reported
in Table 1.

3.3.2 Activity Sample of Twitter Lists

Although the snowball sampling method is convenient and
is easily interpretable with respect to our theoretical motivation, it is also potentially biased by our particular choice
of seeds. To address this concern, we also generated a sample of users based on their activity. Specically, we crawled
5

The blogger category required many more seeds because

bloggers are in general lower prole than the seeds for the
other categories

all lists associated with all users who tweeted at least once
every week for our entire observation period.
This activity-based sampling method is also clearly biased towards users who are consistently active. Importantly,
however, the bias is likely to be quite dierent from any introduced by the snowball sample; despite these dierences,
the qualitative results that follow are similar for both samples, providing evidence that our ndings are not artifacts
of the sampling procedures. This method initially yielded
750k users and 5M lists; however, after pruning the lists to
those that contained at least one of the keywords above, and
assigning users to unique categories (as described above), we
obtained a rened sample of 113,685 users, where Table 1
reports the number of users assigned to each category. We
note that the number of lists obtained by the activity sampling methods is considerably smaller than that obtained
by the snowball sample, and that bloggers are more heavily represented among the activity sample at the expense of
the other three categoriesconsistent with our claim that
the two methods introduce dierent biases. Interestingly,
however, 97,614 of the activity sample, or 85%, also appear
in the snowball sample, suggesting that the two sampling
methods identify similar populations of elite usersas indeed we conrm in the next section.

3.3.3

Classifying Elite Users

Having classied users into the desired categories, we next

rened the categories to identify elite users within each set.
In doing so, we sought to reduce the size of each category
while still accounting for a large fraction of content consumed from these categories. In addition, we xed the four
categories to be of the same size, as categories of very dierent sizes would require us to draw two sets of comparisons
one on the basis of total activity/impact, the other on a
per-capita basisrather than just one. To this end, we rst
ranked all users in each of category by how frequently they
are listed in that category. Next, we measured the ow of information from the top k users in each of the four categories
to a random sample of 100K ordinary (i.e. unclassied) users

50

tweets received

1000

average %
20 30 40
10

4000 7000
top k

10000

1000

4000 7000
top k

10000

(a) Snowball sample

Table 3: Top 5 users in each category

Celebrity
Media
Org
Blog
aplusk
cnnbrk
google
mashable
ladygaga
nytimes
Starbucks problogger
TheEllenShow
asahi
twitter
kibeloco
taylorswift13
BreakingNews
joinred
naosalvo
Oprah
TIME
ollehkt
dooce

50

tweets received
celeb
media
org
blog

10

average %
20 30 40

celeb
media
org
blog

10

average %
20 30 40

50

friends

1000

Table 2: # of URLs initiated by category

# of URLs
category
# of URLs
per-capita
celeb
139,058
27.81
media
5,119,739
1023.94
org
523,698
104.74
blog
1,360,131
272.03
ordinary 244,228,364
6.10

celeb
media
org
blog

10

average %
20 30 40

50

friends
celeb
media
org
blog

4000 7000
top k

10000

1000

4000 7000
top k

10000

(b) Activity sample

Figure 2: Average fraction of # following (blue line)
and # tweets (red line) for a random user that are
accounted for by the top K elites users crawled

in two ways: the proportion of accounts the user follows in

each category, and the proportion of tweets the user received
from everyone the user follows in each category.
Figures 2(a) and 2(b) show the fraction of following links
(square symbols) and tweets received (diamonds) by an average user from each category, respectively. Although the
numerical values dier slightly, the two sets of results are
qualitatively similar. In particular, for both sampling methods, celebrities outrank all other categories, followed by the
media, organizations, and bloggers. Also in both cases, the
bulk of the attention is accounted for by a relatively small
number of users within each category, as evidenced by the
relatively at slope of the attention curves, where we note
that the curve for celebrities asymptotes more slowly than
for the other three categories. Balancing the requirements
described above, therefore, we chose k = 5000 as a cut-o
for the elite categories, where all remaining users are henceforth classied as ordinary. Naturally, imposing categorical
distinctions of any kind articially transforms dierences of
degree (e.g. more or less prominent users) into dierences
of kind (elite vs. ordinary), but again we feel the interpretability gained by this distinction outweighs the costs.
Moreover, because the choice of k = 5000 is arbitrary, we
replicated our analysis with a range of values of k, nding
qualitatively indistinguishable results. Thus, from this point
on, we restrict our analysis to the top 5,000 users in each
category identied by the snowball sampling method, noting
that both methods generate similar results.
Based on this denition of elite users, Table 2 shows that
although ordinary users collectively introduce by far the
highest number of URLs, members of the elite categories are
far more active on a per-capita basis. In particular, users
classied as media easily outproduce all other categories,

followed by bloggers, organizations, and celebrities. Ordinary users originate on average only about 6 URLs each,
compared with over 1,000 for media users. In the rest of
this paper, therefore, when we talk about celebrity, media, organization, blog, we refer the top 5K users drawn
from the snowball sample listed as celebrity, media, organization, blog, respectively.
Table 3, which shows the top 5 users in each of the four
categories, suggests that the sampling method yields results that are consistent with our objective of identifying
users who are prominent exemplars of our target categories.
Among the celebrity list, for example, aplusk, is the handle for actor Ashton Kusher, one of the rst celebrities to
embrace Twitter and still one of the most followed users,
while the remaining celebrity usersLady Gaga, Ellen Degeneres, Oprah Winfrey, and Taylor Swift, are all household
names. In the media category, CNN Breaking News and the
New York Times are most prominent, followed by Breaking
News, Time, and Asahi, a leading Japanese daily newspaper. Among organizations, Google, Starbucks, and Twitter are obviously large and socially prominent corporations,
while JoinRed is the charity organization started by Bono of
U2, and ollehkt is the Twitter account for KT, formerly Korean Telecom. Finally, among the blogging category, Mashable and ProBlogger are both prominent US blogging sites,
while Kibe Loco and Nao Salvo are popular blogs in Brazil,
and dooce is the blog of Heather Armstrong, a widely read
mommy blogger with over 1.5M followers.

4.

WHO LISTENS TO WHOM

The results of the previous section provide qualied support for the conventional wisdom that audiences have become increasingly fragmented. Clearly, ordinary users on
Twitter are receiving their information from many thousands of distinct sources, most of which are not traditional
media organizationseven though media outlets are by far
the most active users on Twitter, only about 15% of tweets
received by ordinary users are received directly from the
media. Equally interesting, however, is that in spite of this
fragmentation, it remains the case that 20K elite users, comprising less than 0.05% of the user population, attract almost
50% of all attention within Twitter. Thus, while attention
that was formerly restricted to mass media channels is now

Category of Twitter Users

Category of Twitter Users

A

Celeb

Org

Media

Blog

A retweet B

B receive tweets from A

Celeb

% of tweets received from

Celeb Media Org Blog
Celeb 38.27 6.23 1.55 3.98
Media 3.91 26.22 1.66 5.69
Org
4.64 6.41 8.05 8.70
Blog
4.94 3.89 1.58 22.55

Org

Media

Blog

Celeb
Celeb 4,334
Media 4,624
Org 1,570
Blog 3,710

# of retweets by
Media
Org Blog
1,489 1,543 5,039
40,263 7,628 32,027
2,539 18,937 11,175
6,382 5,762 99,818

Figure 3: Share of tweets received among elite categories

Figure 4: RT behavior among elite categories

shared amongst other elites, information ows have not

become egalitarian by any means.
The prominence of elite users also raises the question of
how these dierent categories listen to each other. To address this issue, we compute the volume of tweets exchanged
between elite categories. Specically, Figure 3 shows the
average percentage of tweets that category i receives from
category j (indicated by edge thickness), exhibiting noticeable homophily with respect to attention: celebrities overwhelmingly pay attention to other celebrities, media actors
pay attention to other media actors, and so on. The one
slight exception to this rule is that organizations pay more
attention to bloggers than to themselves. In general, in fact,
attention paid by organizations is more evenly distributed
across categories than for any other category.
Figure 3, it should be noted, shows only how many URLs
are received by category i from category j, a particularly
weak measure of attention for the simple reason that many
tweets go unread. A stronger measure of attention, therefore, is to consider instead only those URLs introduced by
category i that are subsequently retweeted by category j.
Figure 4 shows how much information originating from each
category is retweeted by other categories. As with our previous measure of attention, retweeting is strongly homophilous
among elite categories; however, bloggers are disproportionately responsible for retweeting URLs originated by all categories, issuing 93 retweets per person, compared to only 1.1
retweets per person for ordinary users. This result therefore
reects the conventional characterization of bloggers as recyclers and lters of information. Interestingly, however, we
also note that the total number of URLs retweeted by bloggers (465k) is vastly outweighed by the number retweeted by
ordinary users (46M); thus in spite of the much greater percapita activity, their overall impact is still relatively small.

aries? In addition, we may inquire whether these intermediaries, to the extent they exist, are drawn from other elite
categories or from ordinary users, as claimed by the twostep ow theory; and if the latter, in what respects they
dier from other ordinary users.
Before proceeding with this analysis, we note that there
are two ways information can pass through an intermediary
in Twitter. The rst is via retweeting, which occurs when
a users explicitly rebroadcasts a URL that he or she has received from a friend, along with an explicit acknowledgement
of the sourceeither using the ocial retweet functionality
provided by Twitter or by making use of an informal convention such as RT @user or via @user. Alternatively,
a user may tweet a URL that has previously been posted,
but without acknowledgement of a source; in this case we
assume the information was independently rediscovered and
label this a reintroduction of content. For the purposes
of studying when a user receives information directly from
the media or indirectly through an intermediary, we treat
retweets and reintroductions equivalently. If the rst occurrence of a URL in Twitter came from a media user, but a
user received the URL from another source, then that source
can be considered an intermediary, whether they are citing
the source within Twitter by retweeting the URL, or reintroducing it, having discovered the URL outside of Twitter.
To quantify the extent to which ordinary users get their
information indirectly versus directly from the media, we
sampled 1M random ordinary users6 , and for each user,
counted the number n of bit.ly URLs they had received that
had originated from one of our 5K media users, where of
the 1M total, 600K had received at least one such URL.
For each member of this 600K subset we then counted the
number n2 of these URLs that they received via non-media
friends; that is, via a two-step ow. The average fraction
n2 /n = 0.46 therefore represents the proportion of mediaoriginated content that reaches the masses via an intermediary rather than directly. As Figure 5 shows, however,
this average is somewhat misleading. In reality, the population comprises two typesthose who receive essentially
all of their media-originating information via two-step ows
and those who receive virtually all of it directly from the media. Unsurprisingly, the former type is exposed to less total
media than the latter. What is surprising, however, is that

4.1 Two-Step Flow of Information

Examining information ow on Twitter also sheds new
light on the theory of the two-step ow [8], arguably the theory that has most successfully captured the dueling importance of mass media and interpersonal inuence. As we have
already noted, on Twitter the ow of information from the
media to the masses accounts for only a fraction of the total
volume of information. Nevertheless, it is still a substantial
fraction, so it is still interesting to ask: for the special case
of information originating from media sources, what proportion is broadcast directly to the masses, and what proportion
is transmitted indirectly via some population of intermedi-

As before, performing this analysis for the entire population

of over 40M ordinary users proved to be computationally
unfeasible.

random sample
b

# users
0 50000
150000

105

10 102 103 104 105 106

# mediaoriginated URLs

10 102 103 104 105 106

# mediaoriginated URLs

intermediaries

intermediaries
c

# of opinion leaders

104
103
102
10
0

# users
100000

10

102

103

104

105

# of twostep recipients

indirect flow ratio

0.0
0.4
0.8

indirect flow ratio

0.0
0.4
0.8

random sample

10 102 103 104 105 106

# mediaoriginated URLs

10 102 103 104 105 106

# mediaoriginated URLs

Figure 5: Percentage of information that received

via an intermediary as a function of total volume of
media content to which a user is exposed

even users who received up to 100 media URLs during our

observation period received all of them via intermediaries.
Who are these intermediaries, and how many of them are
there? In total, the population of intermediaries is smaller
than that of the users who rely on them, but still surprisingly
large, roughly 490K, the vast majority of which (484K, or
99%) are classied as ordinary users, not elites. To illustrate
the dierence, we note that whereas the top 20K elite users
collectively account for nearly 50% of attention, the top 10K
most-followed ordinary users account for only 5%. Moreover,
Figure 5c also shows that at least some intermediaries also
receive the bulk of their media content indirectly, just like
other ordinary users.
Comparing Figure 5a and 5c, however, we note that intermediaries are not like other ordinary users in that they
are exposed to considerably more media than randomly selected users (9165 media-originated URLs on average vs.
1377), hence the number of intermediaries who rely on twostep ows is smaller than for random users. In addition,
we nd that on average intermediaries have more followers
than randomly sampled users (543 followers versus 34) and
are also more active (180 tweets on average, versus 7). Finally, Figure 6 shows that although all intermediaries, by
denition, pass along media content to at least one other
user, a minority satises this function for multiple users,
where we note that the most prominent intermediaries are
disproportionately drawn from the 4% of elite usersAshton
Kucher (aplusk), for example, acts as an intermediary for
over 100,000 users.
Interestingly, these results are all broadly consistent with
the original conception of the two-step ow, advanced over
50 years ago, which emphasized that opinion leaders were
distributed in all occupational groups, and on every social and economic level, corresponding to our classication
of most intermediaries as ordinary [9]. The original theory
also emphasized that opinion leaders, like their followers,
also received at least some of their information via two-step
ows, but that in general they were more exposed to the
media than their followersjust as we nd here. Finally,
the theory predicted that opinion leadership was not a binary attribute, but rather a continuously varying one, cor-

Figure 6: Frequency of intermediaries binned by #

randomly sampled users to whom they transmit media content
responding to our nding that intermediaries vary widely in
the number of users for whom they act as lters and transmitters of media content. Given the length of time that has
elapsed since the theory of the two-step ow was articulated,
and the transformational changes that have taken place in
communications technology in the interimgiven, in fact,
that a service like Twitter was likely unimaginable at the
timeit is remarkable how well the theory agrees with our
observations.

5.

WHO LISTENS TO WHAT?

The results in Section 4 demonstrate the elite users account for a substantial portion of all of the attention on
Twitter, but also show clear dierences in how the attention
is allocated to the dierent elite categories. It is therefore
interesting to consider what kinds of content is being shared
by these categories. Given the large number of URLs in our
observation period (260M ), and the many dierent ways one
can classify content (video vs. text, news vs. entertainment,
political news vs. sports news, etc.), classifying even a small
fraction of URLs according to content is an onerous task.
Bakshy et al. [1], for example, used Amazons Mechanical
Turk to classify a stratied sample of 1,000 URLs along a
variety of dimensions; however, this method does not scale
well to larger sample sizes.
Instead, we restricted attention to URLs originated by the
New York Times which, with over 2.5M followers, is the most
active and the second-most-followed news organization on
Twitter (after CNN Breaking News). To classify NY Times
content, we exploited a convenient feature of their format
namely that all NY Times URLs are classied in a consistent
way by the section in which they appear (e.g. U.S., World,
Sports, Science, Arts, etc.) 7 . Of the 6398 New York Times
bit.ly URLs we observed, 6370 could be successfully unshortened and assigned to one of 21 categories. Of these, however, only 9 categories had more than 100 URLs during the
observation period, one of whichNY regionwas highly
specic to the New York metropolitan area; thus we focused
our attention on the remaining 8 topical categories. Figure
7 shows the proportion of URLs from each New York Times
section retweeted or reintroduced by each category. World
7
http://www.nytimes.com/year/month/day/category/
title.html?ref=category

1. World News

first observation
of URL

2. U.S. News

0.35

last observation
of URL

0.30
0.25
0.20
0.15
0.10

estimation period = 133 days

0.05

evaluation period = 90 days

0.00
3. Business

4. Sports

0.35
0.30

Total observation window = 223 days

0.25

% RTs and Re-introductions

0.20
0.15

Figure 8: (a) Denition of URL lifespan (b)

Schematic of lifespan estimation procedure

0.10
0.05
0.00
5. Health

6. Technology

0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
7. Science

8. Arts

0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
blog

celeb

media

org

other

blog

celeb

media

org

other

User Category

Figure 7: Number of RTs and Reintroductions of

New York Times stories by content category

news is the most popular category, followed by U.S. News,

Business, and Sports, where increasingly niche categories
like Health, Arts, Science, and Technology are less popular still. In general, the overall pattern is replicated for all
categories of users, but there are some minor deviations: in
particular, organizations show disproportionately little interest in business and arts-related stories, and disproportionately high interest in science, technology, and possibly
world news. Celebrities, by contrast, show greater interest
in sports and less interest in health, while the media shows
somewhat greater interest in U.S. news stories.

5.1 Lifespan of Content

In addition to dierent types of content, URLs introduced
by dierent types of elite users or ordinary users may exhibit
dierent lifespans, by which we mean the time lag between
the rst and last appearance of a given URL on Twitter.
Naively, measuring lifespan seems a trivial matter; however, a nite observation periodwhich results in censoring
of our datacomplicates this task. In other words, a URL
that is last observed towards the end of the observation period may be retweeted or reintroduced after the period ends,
while correspondingly, a URL that is rst observed toward
the beginning of the observation window may in fact have
been introduced before the window began. What we observe
as the lifespan of a URL, therefore, is in reality a lower bound
on the lifespan. Although this limitation does not create
much of a problem for short-lived URLswhich account for
the vast majority of our observationsit does potentially
create large biases for long lived URLs. In particular, URLs
that appear towards the end of our observation period will

be systematically classied as shorter-lived than URLs that

appear towards the beginning.
To address the censoring problem, we seek to determine
a buer at both the beginning and the end of our 223day period, and only count URLs as having a lifespan of
if (a) they do not appear in the rst days, (b) they rst
appear in the interval between the buers, and (c) they do
not appear in the last days, as illustrated in Figure 8(a).
To determine we rst split the 223 day period into two
segmentsthe rst 133 day estimation period and the last
90 day evaluation period (see Figure 8(b))and then ask: if
we (a) observe a URL rst appear in the rst (133 ) days
and (b) do not see it in the days prior to the onset of the
evaluation period, how likely are we see it in the last 90 days?
Clearly this depends on the actual lifespan of the URL, as
the longer a URL lives, the more likely it will re-appear
in the future. Using this estimation/evaluation split, we
nd an upper-bound on lifespan for which we can determine
the actual lifespan with 95% accuracy as a function of .
Finally, because we require a beginning and ending buer,
and because we can only classify a URL as having lifespan
if it appears at least days before the end of our window, we
need to pick and such that + 2 223. We determined
that = 70 and = 70 suciently satised our constraints;
thus for the following analysis, we consider only URLs that
have a lifespan 70 8 .

5.2

Lifespan By Category

Having established a method for estimating URL lifespan,

we now explore the lifespan of URLs introduced by dierent
categories of users, as shown in Figure 9(a). URLs initiated by the elite categories exhibit a similar distribution over
lifespan to those initiated by ordinary users. As Figure 9(b)
shows, however, when looking at the percentage of URLs of
dierent lifespans initiated by each category, we see two additional results: rst, URLs originated by media actors generate a large portion of short-lived URLs (especially URLs
with = 0, those that only appeared once); and second,
URLs originated by bloggers are overrepresented among the
longer-lived content. Both of these results can be explained
by the type of content that originates from dierent sources:
whereas news stories tend to be replaced by updates on a
daily or more frequent basis, the sorts of URLs that are
picked up by bloggers are of more persistent interest, and
so are more likely to be retweeted or reintroduced months
8

We also performed our analysis with dierent values of ,

nding very similar results; thus our conclusions are robust
with respect to the details of our estimation procedure.

108
106
# of URLs

youtube.com
last.fm
amazon.com
pollpigeon.com
en.wikipedia.org
bitrebels.com
mashable.com
feedproxy.google.com
ted.com
ecademy.com
imdb.com
myspace.com
facebook.com
twitter.com
google.com
flickr.com
collegehumor.com
vimeo.com
smashingmagazine.com
123greetings.com

other
celeb
media
org
blog

104
102
0
0

10

20

30

40

50

60

70

lifespan (day)

5
4
3
2
1
0
0

10

20

30
40
50
lifespan (day)

104

Figure 10: Top 20 domains for URLs that lived more

than 200 days

# of RTs
total # of occurrences

celeb
media
org
blog

103
count

60

70

RT rate =

% of URLs from elites category

102

10

(a) Count

(b) Percent
Figure 9: 9(a) Count and 9(b) percentage of URLs
initiated by 4 categories, with dierent lifespans

1.0

other
celeb
media
org
blog

0.8
0.6
0.4
0.2
0.0
0

10

20

30

40

50

60

70

lifespan (day)

or even years after their initial introduction. Twitter, in

other words, should be viewed as a subset of a much larger
media ecosystem in which content exists and is repeatedly
rediscovered by Twitter users. Some of this contentsuch
as daily news storieshas a relatively short period of relevance, after which a given story is unlikely to be reintroduced or rebroadcast. At the other extreme, classic music
videos, movie clips, and long-format magazine articles have
lifespans that are eectively unbounded, and can seemingly
be rediscovered by Twitter users indenitely without losing
relevance.
To shed more light on the nature of long-lived content on
Twitter, we used the bit.ly API service to unshorten 35K
of the most long-lived URLs (URLs that lived at least 200
days), and mapped them into 21034 web domains. As Figure
10 shows, the population of long-lived URLs is dominated
by videos, music, and consumer goods. Two related points
are illustrated by Figure 11, which shows the average RT
rate (the proportion of tweets containing the URL that are
retweets of another tweet) of URLs with dierent lifespans,
grouped by the categories that introduced the URL9 . First,
for ordinary users, the majority of appearances of URLs after the initial introduction derives not from retweeting, but
rather from reintroduction, where this result is especially
pronounced for long-lived URLs. For the vast majority of
URLs on Twitter, in other words, longevity is determined
not by diusion, but by many dierent users independently
9

Note here that URLs with lifespan = 0 are those URLs

that only appeared once in our dataset, thus the RT rate is
zero.

Figure 11: Average RT rate by lifespan for each of

the originating categories

rediscovering the same content, consistent with our interpretation above. Second, however, for URLs introduced by
elite users, the result is somewhat the oppositethat is, they
are more likely to be retweeted than reintroduced, even for
URLs that persist for weeks. Although it is unsurprising
that elite users generate more retweets than ordinary users,
the size of the dierence is nevertheless striking, and suggests that in spite of the dominant result above that content
lifespan is determined to a large extent by the type of content, the source of its origin also impacts its persistence, at
least on averagea result that is consistent with previous
ndings [1].

6.

CONCLUSIONS

In this paper, we investigated a classic problem in media communications research, captured by the rst part of
Laswells maximwho says what to whomin the context
of Twitter. In particular, we nd that although audience attention has indeed fragmented among a wider pool of content
producers than classical models of mass media, attention remains highly concentrated, where roughly 0.05% of the population accounts for almost half of all posted URLs. Within
this population of elite users, moreover, we nd that attention is highly homophilous, with celebrities following celebri-

ties, media following media, and bloggers following bloggers.

Second, we nd considerable support for the two-step ow
of informationalmost half the information that originates
from the media passes to the masses indirectly via a diuse
intermediate layer of opinion leaders, who although classied
as ordinary users, are more connected and more exposed to
the media than their followers. Third, we nd that although
all categories devote a roughly similar fraction of their attention to dierent categories of news (World, U.S., Business,
etc), there are some dierencesorganizations, for example, devote a surprisingly small fraction of their attention
to business-related news. We also nd that dierent types
of content exhibit very dierent lifespans: media-originated
URLs are disproportionately represented among short-lived
URLs while those originated by bloggers tend to be overrepresented among long-lived URLs. Finally, we nd that
the longest-lived URLs are dominated by content such as
videos and music, which are continually being rediscovered
by Twitter users and appear to persist indenitely.
By restricting our attention to URLs shared on Twitter,
our conclusions are necessarily limited to one narrow crosssection of the media landscape. An interesting direction
for future work would therefore be to apply similar methods to quantifying information ow via more traditional
channels, such as TV and radio on the one hand, and interpersonal interactions on the other hand. Moreover, although our approach of dening a limited set of predetermined user-categories allowed for relatively convenient analysis and straightforward interpretation, it would be interesting to explore automatic classication schemes from which
additional user categories could emerge. Finally, another
two areas for future work are rst, to extract content information in a more systematic mannerthe what of Lasswells maxim; and second, to focus more on the eects of
communication by merging the data regarding information
ow on Twitter with other sources of outcome data, such as
the opinions or actions of the recipients of the information.

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]
[16]

[17]

7. REFERENCES
[1] E. Bakshy, J. M. Hofman, W. A. Mason, and D. J.
Watts. Identifying inuencers on twitter. In Fourth
ACM International Conference on Web Seach and
Data Mining (WSDM), Hong Kong, 2011. ACM.
[2] W. L. Bennett and S. Iyengar. A new era of minimal
eects? the changing foundations of political
communication. Journal of Communication,
58(4):707731, 2008.
[3] M. Cha, H. Haddadi, F. Benevenuto, and K. P.
Gummad. Measuring user inuence on twitter: The
million follower fallacy. In 4th Intl AAAI Conference
on Weblogs and Social Media, Washington, DC, 2010.
[4] J. S. Coleman, E. Katz, and H. Menzel. The diusion
of an innovation among physicians. Sociometry,
20(4):253270, 1957.
[5] R. Crane and D. Sornette. Robust dynamic classes
revealed by measuring the response function of a
social system. Proceedings of the National Academy of
Sciences, 105(41):15649, 2008.
[6] T. Gitlin. Media sociology: The dominant paradigm.
Theory and Society, 6(2):205253, 1978.
[7] M. Gomez Rodriguez, J. Leskovec, and A. Krause.
Inferring networks of diusion and inuence. In

Proceedings of the 16th ACM SIGKDD international

conference on Knowledge discovery and data mining,
pages 10191028. ACM, 2010.
E. Katz. The two-step ow of communication: An
up-to-date report on an hypothesis. Public Opinion
Quarterly, 21(1):6178, 1957.
E. Katz and P. F. Lazarsfeld. Personal influence; the
part played by people in the flow of mass
communications. Free Press, Glencoe, Ill. 1955.

G. Kossinets and D. J. Watts. Empirical analysis of an

evolving social network. Science, 311(5757):8890,
2006.
H. Kwak, C. Lee, H. Park, and S. Moon. What is
twitter, a social network or a news media? In
Proceedings of the 19th international conference on
World Wide Web, pages 591600. ACM, 2010.
H. D. Lasswell. The structure and function of
communication in society. In L. Bryson, editor, The
Communication of Ideas, pages 117130. University of
Illinois Press, Urbana, IL, 1948.
P. F. Lazarsfeld, B. Berelson, and H. Gaudet. The
peoples choice; how the voter makes up his mind in a
presidential campaign. Columbia University Press,
New York, 3rd edition, 1968.
R. K. Merton. Patterns of inuence: Local and
cosmopolitan inuentials. In R. K. Merton, editor,
Social theory and social structure, pages 441474. Free
Press, New York, 1968.
C. Sunstein. Going to extremes: how like minds unite
and divide. Oxford University Press, USA, 2009.
J. B. Walther, C. T. Carr, S. S. W. Choi, D. C.
DeAndrea, J. Kim, S. T. Tong, and B. Van Der Heide.
Interaction of interpersonal, peer, and media inuence
sources online. In Z. Papacharissi, editor, A Networked
Self: Identity, Community, and Culture on Social
Network Sites, pages 1738. Routledge, 2010.
J. Weng, E. P. Lim, J. Jiang, and Q. He. Twitterrank:
nding topic-sensitive inuential twitterers. In
Proceedings of the third ACM international conference
on Web search and data mining, pages 261270. ACM,
2010.